License clarification for SenseVoice-Small weights at a fixed revision (ONNX redistribution)
Hello FunASR maintainers,
I am trying to establish what licence governs the SenseVoice-Small weights. This is a licensing question only — no technical support needed.
The artefacts are pinned below so that an answer can be verifiable rather than general.
The artefacts
| ONNX redistribution | csukuangfj/sherpa-onnx-sense-voice-zh-en-ja-ko-yue-2024-07-17 |
| HF revision | 2365baeacb507f821a0c8120fcee3d484dba7a07 |
model.int8.onnx SHA-256 |
c71f0ce00bec95b07744e116345e33d8cbbe08cef896382cf907bf4b51a2cd51 |
tokens.txt SHA-256 |
f449eb28dc567533d7fa59be34e2abca8784f771850c78a47fb731a31429a1dc |
| Upstream weights | FunAudioLLM/SenseVoiceSmall, revision 3847d57b6bdf2dd8875cb1508d2af43d80a16bf7 |
Why the current metadata does not settle it
- The model card at that revision declares
license: other,license_name: model-license, andlicense_link: https://github.com/modelscope/FunASR/blob/main/MODEL_LICENSE. That link points at a branch, not a commit, so it does not fix which text applies. - The ONNX redistribution above ships a
LICENSEwhose entire content isRef to https://github.com/modelscope/FunASR?tab=readme-ov-file#license— a pointer to a README section, not to a licence file or a commit. MODEL_LICENSEis versioned (currently v1.1) and amends itself: "The revised agreement will be published in the official repository … and will take effect automatically." With an unpinned link, a downstream user cannot establish which version they accepted.- §2.1 grants "free to use, copy, modify, and share" with no commercial carve-out, while the following paragraph says the software "is provided for reference and learning purposes only". I cannot tell from the text whether that phrase limits the scope of the grant or is part of the liability disclaimer that surrounds it.
Questions
- For the pinned revision above, is
MODEL_LICENSEv1.1 the applicable licence for the weight files, and does it covermodel.int8.onnxandtokens.txtas listed? - Does "provided for reference and learning purposes only" limit permitted use, or is it part of the disclaimer of warranty? Specifically: is commercial use permitted?
- Does the agreement permit redistributing the weight files — for example bundled inside software delivered to end users — and if so, what satisfies §2.2's "retain relevant model names"?
- Is running the model as a service that returns only recognition results, never the weights, treated differently from redistributing them?
- Are format conversion and quantisation derivative works under this agreement, and are they permitted?
- Does the automatic-amendment clause apply retroactively to a copy obtained under v1.1, or do the terms accepted at download time continue to govern that copy?
- Are there any obligations or attribution requirements arising from the training data
that are not stated in
MODEL_LICENSE?
A suggested resolution
#3454 was resolved by cutting a new
immutable revision containing an actual licence file, tagged apache-2.0-20260804, rather
than by rewriting history. The same approach would work here: a pinned revision of the
SenseVoice-Small weights that carries the applicable licence text would let downstream
users verify their position without having to ask.
#3458 raises the same structural question for four other models and is still open.
Thank you for the clear position taken in #3454 — that the code repository's licence does not automatically cover model weights — which is what prompted me to ask precisely rather than assume.
Source: modelscope/FunASR